Automated safety risk management guidance enhanced by retrieval-augmented large language model

This paper introduces an automated framework for generating safety risk management guidance using a Large Language Model (LLM) enhanced by Retrieval-Augmented Generation (RAG). Reference documents related to specific work activities and equipment are retrieved from 64,740 construction accident cases, generating tailored safety risk management guidance using LLM. This study confirmed that domain adaptation of a text embedding model improves the quality of text retrieval. The generated safety risk management guidance was found to be of equivalent or superior quality to those written by experienced practitioners through a double-blind peer review. In addition, natural language generation (NLG) metrics confirmed the effectiveness of the proposed RAG framework in real-world applications. The findings demonstrate the proposed method to improve safety risk management in construction, making safety practices more consistent and accessible, even for less experienced supervisors.